Write a skill

Last updated: October 12, 2026

Write down how your studio does something, such as its icon rules or its export spec, as a skill, so the Layer agent follows it in every chat where it applies instead of you repeating it in every brief.

Before you start

  • You need a role that can edit skills: Admin, Editor, Art Director or Artist, or a custom role with Edit skills.
  • Decide what the skill is for in one sentence before you start. If it takes two, it is probably two skills. See Skills for how skills load and where they apply.

Steps

  1. Select Skills in the sidebar.

  2. Select Create skill. Under Name, enter a short lowercase handle with dashes, such as ui-icon-style. Under Description, say when the skill should be used, such as "Use when making game UI icons, buttons or HUD elements for Ember Tactics."

  3. Under Loading, choose when the skill reaches the agent: Always, When relevant or Manual. Leave Applies to with every agent ticked unless the rules only fit one kind of work. Select Create skill. The skill's page opens.

  4. Select Edit and write the instructions under Instructions, in markdown: the rules, the steps, what to avoid.

  5. Select Save skill.

  6. To have the skill apply in every session, turn on Enable in whole workspace. To have it apply to one project only, leave it off and add it to the project instead: see Choose where a skill applies.

Result

The skill is in your workspace's Skills library, where the rest of the team can find, use and edit it. Where it is in scope, the agent applies it from the next message anyone sends: a When relevant skill shows as Loaded skill in the reply when it matched, and an Always skill is listed under Preloaded skills at the top of the reply. Any skill can also be run by hand with / and its name in the chat.

Add files to a skill

Under Files on the skill's page, add material the agent opens only when it needs it, so it costs nothing until it is used.

  • references/: select Add, give the file a name such as factions.md, write its contents and select Add reference. Use it for detail too long for the instructions, such as a lore bible or a per-platform export table.
  • assets/: select Add and pick images, video or audio from your library, such as a palette swatch, a logo or a finished example. The agent can pass an asset to a generation as a reference.

Then point to them from the instructions, so the agent knows when to open them: "For a faction's colours, read references/factions.md." Select Copy path for the instructions on a file to copy its path.

Have the agent write it

You can also ask the agent in any session to write or change a skill for you, for example:

  • "Turn the icon rules we agreed in this chat into a skill and keep to it from now on."
  • "Update my sprite-export skill: we use 512 px frames now, not 256."

The agent confirms the name and the change before it saves, and can add the new skill to the session's project or to a reference set you name. It cannot turn a skill on for the whole workspace: do that on the skill's page.

Production tips

  • Write the description as a trigger ("Use when exporting sprite frames for the build"), not a title. With When relevant, the description is all the agent reads before deciding to load the skill.
  • Give numbers, not adjectives: "512 x 512, 6 px outline, readable at 64 px" is followed; "bold, clean icons" is interpreted.
  • Keep the instructions to about a page. They go into the agent's context in full whenever the skill applies, so move long detail into a reference file.
  • Use Always sparingly. It loads the skill on every message, including unrelated ones; save it for rules that truly apply to everything, such as brand and content rules.
  • Narrow Applies to for specialist rules. A Spine rig naming convention needs only Spine animation; ad rules need Marketing video, UA expert and Playable generation.
  • Test the skill in a fresh session with a request it should cover, and check that the reply shows it. For ideas and copyable instructions, see Skill examples for game studios.

If something goes wrong

  • "Use lowercase letters, digits and single dashes between them". The name must be lowercase words joined by single dashes, such as pixel-art, and unique in the workspace.
  • The skill never loads on its own. Check that it has a description, that its loading mode is not Manual, that the agent handling the chat is ticked under Applies to, and that the skill is in scope: enabled for the workspace, added to the session's project, on its board or attached to a reference set in use. Then make the description name the task more directly.
  • The instructions are over the limit. Move a section into a reference file under Files and tell the agent in the instructions when to read it.
  • You renamed a skill and an old instruction stopped working. The name is the handle agents and people use to call the skill. Update anything that refers to the old name, such as another skill's instructions.

Related